A standardised differential privacy framework for epidemiological modelling with mobile phone data

Merveille Koissi Savi, Akash Yadav, Wanrong Zhang, Navin Vembar, Andrew Schroeder, Satchit Balsari, Caroline O. Buckee, Salil Vadhan, Nishant KishoremedRxiv 2023 · DOI

Abstract

During the COVID-19 pandemic, the use of mobile phone data for monitoring human mobility patterns became a cornerstone of epidemiological modelling and public policy. However, the use of granular location information raises significant ethical concerns regarding individual privacy. While simple aggregation is often insufficient to guarantee anonymity, newer methods like differential privacy (DP) provide statistically verifiable protection but remain largely untested in the context of infectious disease compartment models. This study introduces a standardised framework to evaluate the impact of DP-induced noise on ten common epidemiological metrics. By applying the Laplace mechanism to mobile phone-based location data, we find that most metrics remain stable and unbiased when the privacy loss parameter is maintained at $\epsilon = 0.05$ per release (corresponding to a noise state of less than 20). Our findings suggest that DP offers a robust approach to balancing individual privacy with the need for population-level public health insights.

Introduction

The tension between data utility and individual privacy has never been more apparent than during the global response to COVID-19. Mobile phone mobility data, often derived from GPS logs, allowed researchers to track the impact of travel restrictions and predict the trajectory of the virus with unprecedented precision. Yet, even aggregated datasets can be de-identified, potentially revealing sensitive information about individuals' movements between home and work.

To address this, researchers have turned to Differential Privacy (DP), a mathematical framework that adds controlled statistical noise to datasets. The challenge lies in the "privacy-utility trade-off": too much noise renders the data useless for public health decisions, while too little noise fails to protect individuals. This paper presents a modular software pipeline designed to quantify this trade-off and establish a standard for privacy-preserving mobility data in epidemiology.

Methodology

The researchers developed a "plug-and-play" pipeline to test the resilience of epidemiological models against DP noise.

The Privacy Mechanism

The framework utilizes the Laplace mechanism, a standard DP technique that adds noise to data counts based on a defined privacy budget ($\epsilon$). To manage the cumulative privacy loss over multiple data releases (e.g., daily mobility updates), the team implemented an advanced composition theorem using the "smartnoise sdk."

Epidemiological Modelling

The noisy mobility data was integrated into a metapopulation model—specifically a Susceptible-Exposed-Infected (Symptomatic/Asymptomatic) framework. This model simulates disease transmission across different geographic "patches" (counties). To ensure the findings were not specific to a single set of conditions, the researchers conducted 1,000 Monte Carlo iterations across eight scenarios, varying parameters such as outbreak location (urban vs. rural), regional connectivity, and transmission rates.

Data Source

The study used real-world mobility data from Camber Systems, covering 62 counties in New York State over a 90-day period (August 15 to November 15, 2020). This dataset included over 812,000 transitions, representing approximately 3-7% of the population. These counts were normalized using American Community Survey (ACS) data before the DP framework was applied.

Key Insights and Results

The $\epsilon = 0.05$ Threshold

The study identified a critical "noise threshold." Most epidemiological metrics—including total epidemic size and peak timing—remained stable and unbiased until the "noise state" exceeded 20. This level of noise corresponds to a privacy loss parameter of $\epsilon = 0.05$ per matrix release. Beyond this threshold, the statistical "noise" begins to drown out the actual "signal" of human movement, leading to unreliable model outputs.

Rural vs. Urban Sensitivity

One of the most significant findings was the differential impact of noise based on geography. Rural and isolated areas are far more sensitive to DP noise than high-traffic urban centers. Because rural routes have fewer total transitions, the relative impact of adding noise is much higher. This can create "phantom" connections in the model, leading to an upward bias in importation risk—the predicted likelihood of a disease entering a previously uninfected area.

Metric Resilience

While some metrics like the "rate of spread" were prone to bias as noise increased, others proved remarkably resilient. The timing of the epidemic peak and the overall size of the outbreak were largely preserved under moderate noise levels, suggesting that DP-protected data is highly suitable for long-term strategic planning, even if real-time granular tracking is slightly obscured.

Discussion and Conclusion

This research provides a much-needed standard for the public health community. By suggesting a quantifiable threshold of $\epsilon = 0.05$, the authors offer a benchmark for data providers and researchers to share mobility data responsibly.

Limitations

The authors note that these results are based on a specific metapopulation model; different architectures, such as agent-based models, may exhibit different sensitivities to noise. Additionally, the raw data used already contained some baseline privacy protections from the provider, which complicates the establishment of a perfectly "clean" non-private baseline.

Future Directions

Future work should investigate alternative DP mechanisms beyond the Laplace distribution and test the framework's performance in real-time forecasting scenarios. By releasing their pipeline as modular software, the authors have enabled the broader research community to replicate these findings and continue refining the balance between individual privacy and the collective good of public health.

Summary updated

Abstract

Abstract During the COVID-19 pandemic, the use of mobile phone data for monitoring human mobility patterns has become increasingly common, both to study the impact of travel restrictions on population movement and epidemiological modelling. Despite the importance of these data, the use of location information to guide public policy can raise issues of privacy and ethical use. Studies have shown that simple aggregation does not protect the privacy of an individual, and there are no universal standards for aggregation that guarantee anonymity. Newer methods, such as differential privacy, can provide statistically verifiable protection against identifiability but have been largely untested as inputs for compartment models used in infectious disease epidemiology. Our study examines the application of differential privacy as an anonymisation tool in epidemiological models, studying the impact of adding quantifiable statistical noise to mobile phone-based location data on the bias of ten common epidemiological metrics. We find that many epidemiological metrics are preserved and remain close to their non-private values when the true noise state is less than 20, in a count transition matrix, which corresponds to a privacy-less parameter ∈ = 0.05 per release. We show that differential privacy offers a robust approach to preserving individual privacy in mobility data while providing useful population-level insights for public health. Importantly, we have built a modular software pipeline to facilitate the replication and expansion of our framework. Author Summary Human mobility data has been used broadly in epidemiological population models to better understand the transmission dynamics of an epidemic, predict its future trajectory, and evaluate potential interventions. The availability and use of these data inherently raises the question of how we can balance individual privacy and the statistical utility of these data. Unfortunately, there are few existing frameworks that allow us to quantify this trade-off. Here, we have developed a framework to implement a differential privacy layer on top of human mobility data which can guarantee a minimum level of privacy protection and evaluate their effects on the statistical utility of model outputs. We show that this set of models and their outputs are resilient to high levels of privacy-preserving noise and suggest a standard privacy threshold with an epsilon of 0.05. Finally, we provide a reproducible framework for public health researchers and data providers to evaluate varying levels of privacy-preserving noise in human mobility data inputs, models, and epidemiological outputs.
BibTeX
@misc{savi2023standardised,
  author = {Merveille Koissi Savi and Akash Yadav and Wanrong Zhang and Navin Vembar and Andrew Schroeder and Satchit Balsari and Caroline O. Buckee and Salil Vadhan and Nishant Kishore},
  title = {{A standardised differential privacy framework for epidemiological modelling with mobile phone data}},
  year = {2023},
  doi = {10.1101/2023.03.16.23287382},
  howpublished = {medRxiv},
}
Figure